ChatGPT is saying wrong things about your company. Here's the factual-correction workflow.

A prospect says it on a sales call: "I asked ChatGPT about you first." And what ChatGPT told them was your 2023 pricing, a feature you sunset last year, or — the special horror — facts belonging to a different company with a similar name. Every day this goes unfixed, an unknown number of buyers are being quietly mis-sold your product by a machine you never hired.

The three causes (and why the fix differs for each)

• Stale training data. The model's snapshot predates your change. It answers "from memory," and its memory is your old website. Fix: get the current fact into the high-authority sources the next training pass will ingest — and into retrievable pages meanwhile.
• A bad retrieved source. With browsing on, ChatGPT pulls live pages — and one of them (an old directory listing, a stale review, a 2024 blog post) contains the wrong fact. Fix: correct or displace that specific page. This is the fastest error class to fix.
• Entity confusion. A similarly named company's facts are bleeding into yours. Fix: disambiguation — explicit "not affiliated with…" statements, consistent Organization schema, and aligned profiles everywhere your name appears.

A quick manual test distinguishes them: ask the same question with web search on, then off. Wrong only with search on → a retrieved page is the culprit. Wrong with search off → stale training data. Wrong in both → likely both problems, or entity confusion.

Worked audit: a real buyer prompt, claim by claim

Sample Factual Check output for the prompt "Is AcmeFlow good for a 20-person marketing team? How does it compare to TaskRiver?" — the kind of comparison prompt buyers actually type. (Illustrative company names; run your own brand prompts for the live version.)

The corrections shipped, one per error class:

• Pricing (training-stale): published on the pricing page as an extractable dated statement — "As of January 2026, AcmeFlow starts at $12/user/month, with annual billing at 15% off." Also synced to G2, Capterra, and Crunchbase, the profiles training passes ingest.
• SSO (training-stale): added to /changelog and the security page — "SSO (SAML 2.0) is included from the Team tier as of February 2026." Changelogs are retrieval magnets; date every entry.
• Acquisition (entity confusion): added to the About page — "AcmeFlow, Inc. is independent and bootstrapped. We are not affiliated with AcmeFlow Analytics." Plus Organization schema with distinct legal name, founding date, and sameAs links on every profile.
• Support (bad retrieved source): published a dated support-metrics page ("median first response: 42 minutes, measured June 2026") — a fresher, more authoritative passage for retrieval to prefer over the 2023 thread.

Re-test at 3 weeks: the retrieval-based errors (acquisition, support) corrected; the pricing figure updated in browse-on mode but persisted in browse-off — now queued against the next model refresh. That split is exactly what the timeline section below predicts.

The correction workflow

1. Audit the buyer prompts. Run the 10–15 prompts that matter commercially — "what is [brand]," "is [brand] worth it," "[brand] pricing," "[brand] vs [each rival]," "best [category] for [your ICP]" — through the ChatGPT Gap Analyzer. Export every flagged issue with its correction and evidence.
2. Classify each error with the search-on/search-off test. Retrieval errors get fixed this week; training errors get queued for authority building.
3. Publish extractable corrections. For each wrong fact, make the true fact exist as a plain, dated, self-contained statement on the page retrieval is most likely to hit — pricing page, changelog, About page. Marketing language hides facts from models; declarative sentences expose them.
4. Align the third-party record. LinkedIn, Crunchbase, G2, and directories must agree with your site. Conflicting sources are how confident hallucinations get manufactured.
5. Re-test on a schedule. Answers are probabilistic — re-run each prompt several times, monthly. When an error survives, the Factual Check shows which claim persists so you can trace which source still feeds it.

Realistic timelines

Set expectations honestly: corrections to retrieved sources typically surface in ChatGPT within days to a few weeks of recrawl. Corrections that depend on training data wait for the next model refresh — historically anywhere from two to six months. The audit-and-publish work is identical either way, which is why the right time to start is before anyone asks.

The three OpenAI crawlers — and why each one changes your fix

Most "optimize for ChatGPT" advice treats OpenAI's crawling as one thing. It's three, with different jobs, and the distinction decides where your corrections must live:

• GPTBot gathers content for training. What it collects shapes what future model versions "remember" about you. Blocking it in robots.txt means future models learn about your brand only from third parties — usually the opposite of what a brand wants.
• OAI-SearchBot builds the search index behind ChatGPT search. If it can't reach your pricing and changelog pages, browse-on answers ground on whatever it can reach — directories, old reviews, competitors.
• ChatGPT-User fetches pages live, at the user's request, when the assistant visits a specific URL mid-conversation. This is your fastest correction path: a fixed page can change an answer the same day it's fetched.

Now re-read the timeline section above with this lens: "training-stale" errors are GPTBot-era snapshots (fix waits for a model refresh), while retrieval errors are OAI-SearchBot/ChatGPT-User problems (fix lands in days). Our site's own robots.txt explicitly welcomes all three — practicing what this page preaches.

Second worked example: the reputation-query audit

Beyond comparison prompts, audit the fear query — "AcmeFlow problems and complaints":

Go deeper: the same passage-level evidence standards that fix ChatGPT answers win Google citations — see what LLMs check before citing you — and if you're weighing monitoring tools for brand prompts, start with the honest category map.

Frequently asked questions

Dmitry Dragilev

4× acquired founder: Polar Polls → Google (2014), JustReachOut → SEOJet (2020), Smallbiz.Tools → Semrush (2023), SERP Gap Analyzer — the #1 top-performing app in the Semrush App Center → Semrush (2025). Contributor to Forbes, TechCrunch, WIRED, and Moz since 2009. More at criminallyprolific.com.